[SYCL] Improve record and replay edge determination

sycl/source/detail/graph_impl.cpp

@@ -54,20 +54,63 @@ void graph_impl::remove_root(node_ptr n) {
   MSchedule.clear();
 }
 
+// Recursive check if a graph node or its successors contains a given kernel
+// argument.
+//
+// @param[in] arg The kernel argument to check for.
+// @param[in] currentNode The current graph node being checked.
+// @param[in,out] deps The unique list of dependencies which have been
+// identified for this arg.
+// @param[in] dereferencePtr if true arg comes direct from the handler in which
+// case it will need to be deferenced to check actual value.
+//
+// @returns True if a dependency was added in this node of any of its
+// successors.
+bool check_for_arg(const sycl::detail::ArgDesc &arg, node_ptr currentNode,
+                   std::set<node_ptr> &deps, bool dereferencePtr = false) {
+  bool successorAddedDep = false;
+  for (auto &successor : currentNode->MSuccessors) {
+    successorAddedDep |= check_for_arg(arg, successor, deps, dereferencePtr);
+  }
+
+  if (deps.find(currentNode) == deps.end() &&
+      currentNode->has_arg(arg, dereferencePtr) && !successorAddedDep) {
+    deps.insert(currentNode);
+    return true;
+  }
+  return successorAddedDep;
+}
+
 template <typename T>
 node_ptr graph_impl::add(graph_ptr impl, T cgf,
+                         const std::vector<sycl::detail::ArgDesc> &args,
                          const std::vector<node_ptr> &dep) {
-  node_ptr nodeImpl = std::make_shared<node_impl>(impl, cgf);
-  if (!dep.empty()) {
-    for (auto n : dep) {
+  node_ptr nodeImpl = std::make_shared<node_impl>(impl, cgf, args);
+  // Copy deps so we can modify them
+  auto deps = dep;
+  // A unique set of dependencies obtained by checking kernel arguments
+  std::set<node_ptr> uniqueDeps;
+  for (auto &arg : args) {
+    if (arg.MType != sycl::detail::kernel_param_kind_t::kind_pointer) {
+      continue;
+    }
+    // Look through the graph for nodes which share this argument
+    for (auto nodePtr : MRoots) {
+      check_for_arg(arg, nodePtr, uniqueDeps, true);
+    }
+  }
+
+  // Add any deps determined from arguments into the dependency list
+  deps.insert(deps.end(), uniqueDeps.begin(), uniqueDeps.end());
+  if (!deps.empty()) {
+    for (auto n : deps) {
       n->register_successor(nodeImpl); // register successor
       this->remove_root(nodeImpl);     // remove receiver from root node
                                        // list
     }
   } else {
     this->add_root(nodeImpl);
   }
-  MLastNode = nodeImpl;
   return nodeImpl;
 }
 
@@ -93,7 +136,7 @@ node command_graph<graph_state::modifiable>::add_impl(
     depImpls.push_back(sycl::detail::getSyclObjImpl(d));
   }
 
-  auto nodeImpl = impl->add(impl, cgf, depImpls);
+  auto nodeImpl = impl->add(impl, cgf, {}, depImpls);
   return sycl::detail::createSyclObjFromImpl<node>(nodeImpl);
 }
 

sycl/source/detail/graph_impl.hpp

@@ -8,6 +8,7 @@
 
 #pragma once
 
+#include <sycl/detail/cg_types.hpp>
 #include <sycl/ext/oneapi/experimental/graph.hpp>
 #include <sycl/handler.hpp>
 
@@ -53,6 +54,8 @@ struct node_impl {
 
   std::function<void(sycl::handler &)> MBody;
 
+  std::vector<sycl::detail::ArgDesc> MArgs;
+
   void exec(sycl::detail::queue_ptr q);
 
   void register_successor(node_ptr n) {
@@ -65,7 +68,17 @@ struct node_impl {
   sycl::event get_event(void) const { return MEvent; }
 
   template <typename T>
-  node_impl(graph_ptr g, T cgf) : MScheduled(false), MGraph(g), MBody(cgf) {}
+  node_impl(graph_ptr g, T cgf, const std::vector<sycl::detail::ArgDesc> &args)
+      : MScheduled(false), MGraph(g), MBody(cgf), MArgs(args) {
+    for (size_t i = 0; i < MArgs.size(); i++) {
+      if (MArgs[i].MType == sycl::detail::kernel_param_kind_t::kind_pointer) {
+        // Make sure we are storing the actual USM pointer for comparison
+        // purposes, note we couldn't actually submit using these copies of the
+        // args if subsequent code expects a void**.
+        MArgs[i].MPtr = *(void **)(MArgs[i].MPtr);
+      }
+    }
+  }
 
   // Recursively adding nodes to execution stack:
   void topology_sort(std::list<node_ptr> &schedule) {
@@ -76,11 +89,23 @@ struct node_impl {
     }
     schedule.push_front(node_ptr(this));
   }
+
+  bool has_arg(const sycl::detail::ArgDesc &arg, bool dereferencePtr = false) {
+    for (auto &nodeArg : MArgs) {
+      if (arg.MType == nodeArg.MType && arg.MSize == nodeArg.MSize) {
+        // Args coming directly from the handler will need to be dereferenced
+        // since they are actually void**
+        void *incomingPtr = dereferencePtr ? *(void **)arg.MPtr : arg.MPtr;
+        if (incomingPtr == nodeArg.MPtr) {
+          return true;
+        }
+      }
+    }
+    return false;
+  }
 };
 
 struct graph_impl {
-  // The last node added to the graph.
-  node_ptr MLastNode;
   std::set<node_ptr> MRoots;
   std::list<node_ptr> MSchedule;
   // TODO: Change one time initialization to per executable object
@@ -95,9 +120,9 @@ struct graph_impl {
   void remove_root(node_ptr n);
 
   template <typename T>
-  node_ptr add(graph_ptr impl, T cgf, const std::vector<node_ptr> &dep = {});
-
-  node_ptr getLastNode() const { return MLastNode; }
+  node_ptr add(graph_ptr impl, T cgf,
+               const std::vector<sycl::detail::ArgDesc> &args,
+               const std::vector<node_ptr> &dep = {});
 
   graph_impl() : MFirst(true) {}
 };

sycl/source/detail/queue_impl.hpp

@@ -537,20 +537,14 @@ class queue_impl {
                     const SubmitPostProcessF *PostProcess) {
     event Event = detail::createSyclObjFromImpl<event>(
         std::make_shared<detail::event_impl>());
+    handler Handler(Self, PrimaryQueue, SecondaryQueue, MHostQueue);
+    Handler.saveCodeLoc(Loc);
+    CGF(Handler);
     if (auto graphImpl = Self->getCommandGraph(); graphImpl != nullptr) {
-
-      // TODO: Simple implementation schedules all recorded nodes in order with
-      // each having a dependency on the previous node. This should be improved
-      // to correctly determine edges and dependencies.
-      std::vector<ext::oneapi::experimental::detail::node_ptr> deps;
-      if (auto lastNode = graphImpl->getLastNode(); lastNode != nullptr) {
-        deps.push_back(lastNode);
-      }
-      graphImpl->add(graphImpl, CGF, deps);
+      // Pass the args obtained by the handler to the graph to use in
+      // determining edges between this node and previously submitted nodes.
+      graphImpl->add(graphImpl, CGF, Handler.MArgs, {});
     } else {
-      handler Handler(Self, PrimaryQueue, SecondaryQueue, MHostQueue);
-      Handler.saveCodeLoc(Loc);
-      CGF(Handler);
 
       // Scheduler will later omit events, that are not required to execute
       // tasks. Host and interop tasks, however, are not submitted to low-level

sycl/test/graph/graph-record-dotp-buffer.cpp

@@ -0,0 +1,112 @@
+// RUN: %clangxx -fsycl -fsycl-targets=%sycl_triple %s -o %t.out
+#include <CL/sycl.hpp>
+#include <iostream>
+#include <thread>
+
+#include <sycl/ext/oneapi/experimental/graph.hpp>
+
+const size_t n = 10;
+
+float host_gold_result() {
+  float alpha = 1.0f;
+  float beta = 2.0f;
+  float gamma = 3.0f;
+
+  float sum = 0.0f;
+
+  for (size_t i = 0; i < n; ++i) {
+    sum += (alpha * 1.0f + beta * 2.0f) * (gamma * 3.0f + beta * 2.0f);
+  }
+
+  return sum;
+}
+
+int main() {
+  float alpha = 1.0f;
+  float beta = 2.0f;
+  float gamma = 3.0f;
+
+  sycl::property_list properties{
+      sycl::property::queue::in_order(),
+      sycl::ext::oneapi::property::queue::lazy_execution{}};
+
+  sycl::queue q{sycl::gpu_selector_v, properties};
+
+  sycl::ext::oneapi::experimental::command_graph g;
+
+  float dotpData = 0.f;
+  std::vector<float> xData(n);
+  std::vector<float> yData(n);
+  std::vector<float> zData(n);
+
+  {
+    sycl::buffer dotpBuf(&dotpData, sycl::range<1>(1));
+
+    sycl::buffer xBuf(xData);
+    sycl::buffer yBuf(yData);
+    sycl::buffer zBuf(zData);
+
+    q.begin_recording(g);
+
+    /* init data on the device */
+    q.submit([&](sycl::handler &h) {
+      auto x = xBuf.get_access(h);
+      auto y = yBuf.get_access(h);
+      auto z = zBuf.get_access(h);
+      h.parallel_for(n, [=](sycl::id<1> it) {
+        const size_t i = it[0];
+        x[i] = 1.0f;
+        y[i] = 2.0f;
+        z[i] = 3.0f;
+      });
+    });
+
+    q.submit([&](sycl::handler &h) {
+      auto x = xBuf.get_access(h);
+      auto y = yBuf.get_access(h);
+      h.parallel_for(sycl::range<1>{n}, [=](sycl::id<1> it) {
+        const size_t i = it[0];
+        x[i] = alpha * x[i] + beta * y[i];
+      });
+    });
+
+    q.submit([&](sycl::handler &h) {
+      auto y = yBuf.get_access(h);
+      auto z = zBuf.get_access(h);
+      h.parallel_for(sycl::range<1>{n}, [=](sycl::id<1> it) {
+        const size_t i = it[0];
+        z[i] = gamma * z[i] + beta * y[i];
+      });
+    });
+
+    q.submit([&](sycl::handler &h) {
+      auto dotp = dotpBuf.get_access(h);
+      auto x = xBuf.get_access(h);
+      auto z = zBuf.get_access(h);
+      h.parallel_for(sycl::range<1>{n}, [=](sycl::id<1> it) {
+        const size_t i = it[0];
+        // Doing a manual reduction here because reduction objects cause issues
+        // with graphs.
+        if (i == 0) {
+          for (size_t j = 0; j < n; j++) {
+            dotp[0] += x[j] * z[j];
+          }
+        }
+      });
+    });
+
+    q.end_recording();
+
+    auto exec_graph = g.finalize(q.get_context());
+
+    q.submit([&](sycl::handler &h) { h.exec_graph(exec_graph); });
+  }
+
+  if (dotpData != host_gold_result()) {
+    std::cout << "Error unexpected result!\n";
+  }
+
+  std::cout << "done.\n";
+
+  return 0;
+}